Welcome!
I graduated in 2020 from the University of Michigan with a bachelor’s in computer engineering. After graduating I worked as an embedded engineer at Apple and later Atomic Technologies. I recently graduated from Boston University with a research-based master’s in ecology, evolution, and behavior.
This website showcases my research experience, personal projects, and interests. If you would like to contact me, feel free to reach out at ritikasibal2@gmail.com!
Wildlife x Technology Work
Orangutan pose estimation
Goal: develop a 2D pose estimation tool to automatically detect joint position in videos of wild orangutans.
Publication detailing the full method coming soon! In the mean time, here are some example videos of the model in action!
Dolphin swimming (IMU)
Goal: Implement a model to predict the behaviors of dolphins with a high degree of certainty based on IMU data collected from dolphins wearing bio-logging tags.
For this case study, we monitored dolphins in an enclosed sanctuary in Hawaii. Data was provided by the Barton Research group and included camera data (overhead and on body) and IMU data. We decide to label 4 behaviors (straight, sideways, upside-down and respiration).
I developed an interactive MATLAB GUI to ease the data collection process and build a valid and accurate training set. Based on camera data I tagged data based on what the dolphins doing.
Once a good training data set was achieved, we developed a hybrid machine learning model (pictured below)
We then trained the model (top) and created a state model (bottom) describing the probability of making state transitions, providing insight into the dolphin’s decision making.
With this data and process, we can actually draw larger biological conclusions such as dolphin personality types!
Dolphin swimming (computer vision)
This was a personal project I embarked upon using the same data and problem statement as above Instead of using IMU data, I used computer vision to draw swimming state conclusions.
I converted the video stream from the onboard camera to single frames at a 5 fps rate. From there, I manually grouped the frames into different categories (pictured below). As you can see here, there are distinct lighting characteristics in each of the states, making it a good candidate for a computer vision model.
I chose a CNN model for my neural network. After categorizing the data, the model was trained and we achieved a 95% accuracy!
